Foundations of Modern AI Systems
LLM mechanics, Transformers, tokens, context windows, hallucinations, and training stages — the base layer for every AI product decision.
Model Landscape, Capabilities & Economics
Model families, multimodal AI, benchmarks, token economics, and the model selection memo — compare providers like infrastructure, not hype.
Full 8-module learning path
Scan by theme, weeks, and deliverable. Modules build sequentially from mechanics to strategy.
Module 01 · Weeks 1–2 · 2 weeks · LLM Mechanics
Foundations of Modern AI Systems
Every token costs money. Understanding tokenization, context windows, and model tiers lets you set product constraints and own AI feature unit economics.
Deliverable: Model selection memo — use case, quality reqs, cost estimate at scale, chosen model + rationale, risks
Module 02 · Weeks 3–4 · 2 weeks · Model Selection & Economics
Model Landscape, Capabilities & Economics
Multimodal, model tiers, benchmarks, cost modelling — compare model families like infrastructure choices, not brand fandom.
Deliverable: One-page model selection memo with eval method, cost estimate, rationale, risks, and monitoring
Module 03 · Weeks 3–4 · 2 weeks · Prompting Craft
Prompt Engineering — Techniques & Structure
Prompt engineering is the new acceptance criteria. A well-structured prompt defines "done" for the AI as clearly as a user story does for an engineer.
Deliverable: 3 production-quality system prompts for features in your product + an eval rubric for each
Module 03 · Weeks 4–5 · 1.5 weeks · Cost Optimization
Token-Efficient Prompting
At 1M calls/day, a 3x reduction in prompt tokens can be the difference between a profitable and unprofitable AI feature. Token efficiency is a product economics lever.
Deliverable: Token efficiency audit — compress 3 prompts by ≥50%, validate quality parity, document savings at scale
Module 04 · Weeks 6–8 · 3 weeks · Knowledge Grounding
Retrieval-Augmented Generation (RAG)
RAG is the most common production pattern for grounding LLMs in company knowledge. Retrieval precision directly reduces context tokens — a well-tuned retriever cuts input size 40–60%.
Deliverable: RAG pipeline spec — chunking strategy, embedding model, vector DB, retrieval approach, eval plan, freshness strategy
Module 05 · Weeks 9–12 · 4 weeks · Autonomous AI
AI Agents & Agentic Systems
Agents introduce compounding failure modes and latency. Your job is defining the right HITL boundaries and what "done" looks like for an agent task.
Deliverable: Agentic feature PRD — scope, tool list, HITL checkpoints, blast radius controls, observability, token cost per run
Module 06 · Weeks 12–14 · 2 weeks · Agent Connectivity
Model Context Protocol (MCP)
MCP is becoming the USB-C of AI tool connectivity. Understanding when to expose your product via MCP is a platform strategy question.
Deliverable: MCP build-vs-skip decision memo — who benefits, tools to expose, security requirements
Module 07 · Weeks 14–17 · 3 weeks · Production AI
AI Infrastructure & the Production Stack
Observability tools let you see token spend by feature, by user segment, and by prompt version — this is your cost-of-goods dashboard for AI.
Deliverable: AI observability spec — metrics, dashboards, alerts, cost-per-feature breakdown
Module 08 · Weeks 17–22 · 5 weeks · Strategy & Capstone
AI Product Strategy & Applied Practice
The capstone synthesizes everything. Pick a real problem and design the full AI system — including token budget, caching strategy, eval plan, guardrails, and product decisions.
Deliverable: Capstone — full agentic feature design with PRD, token budget, eval strategy, guardrails, HITL design, unit economics
Who this is for
AI Learning is a structured book-style path — not scattered posts. Modules build sequentially from mechanics to strategy.
Written for product leaders, PMs, founders, AI implementation owners, and operators who want to understand AI systems deeply enough to build, evaluate, and lead them — without becoming ML researchers.